# Darmstadt's Daisytuner Raises EUR 6M Seed Co-Led by Elaia and LEA Ventures to Make AI Models Run Fast on Any Chip
> Daisytuner raises EUR 6M seed co-led by Elaia and LEA Ventures to scale Daisyflow, a self-learning compiler that tunes AI models for Nvidia, AMD and new chips.
- **Author:** Swati k jagtap
- **Published:** 2026-11-16
- **Updated:** 2026-11-16
- **Category:** Funding
- **Editorial type:** news
- **Topics:** seed-funding, lea-ventures, german-startups
- **Canonical URL:** https://www.todaysstartupnews.com/news/daisytuner-raises-6m-seed-elaia-lea-ventures-ai-compiler-any-chip
## Key takeaways
- German deeptech startup Daisytuner has raised EUR 6 million in seed funding co-led by Elaia and LEA Ventures, with Amadeus APEX Technology Fund and Seedcamp participating
- The company automates performance engineering, tuning AI models for the chips they run on, using a self-learning compiler and a patented optimisation platform
- Its open-source compiler turns code from frameworks such as PyTorch into native code for Nvidia and AMD GPUs and newer architectures
- Daisyflow, its inference engine, is live in production, and the company reports Qwen 2.5 running up to 18 percent faster than vLLM on AMD Instinct GPUs
- Funds will bring Daisyflow to more enterprises, grow the team and expand the compiler technology
Daisytuner, a German deeptech startup based in Darmstadt, has raised EUR 6 million in seed funding to make AI models run efficiently on any chip. Elaia and LEA Ventures co-led the round, with participation from Amadeus APEX Technology Fund and Seedcamp. The company announced the funding on November 16, 2026.

Daisytuner will use the money to bring its Daisyflow platform to more enterprises, grow its team and expand its compiler technology. Daisyflow is its inference engine and is now live in production.

The company was founded in 2024 by Lukas Trümper, who is chief executive, and Adrian Schmitz. Its technology builds on more than a decade of research, and the team comes from research institutions including ETH Zurich, TU Munich, RWTH Aachen, TU Darmstadt and Lawrence Livermore National Laboratory.

Daisytuner automates performance engineering, the specialised work of tuning AI models for the chips they run on. The company argues that this has become a bottleneck. Open-weight models are now strong enough for companies to run advanced AI on their own hardware, and as AI agents multiply the compute each task needs, the cost and speed of AI depend less on the model than on how well it runs on the chip beneath it. New hardware is also arriving, from established GPUs to a wave of new accelerators, each with its own optimisation needs. Performance engineering remains largely manual and is concentrated in a few frontier labs, and every new model or chip can break earlier optimisations. As a result, most hardware runs well below its potential.

Daisytuner replaces that manual work with a self-improving process. Its open-source compiler turns code from frameworks such as PyTorch into native code for each chip, from Nvidia and AMD GPUs to next-generation architectures. The company's product pages list support for Nvidia, AMD, Intel and Tenstorrent hardware. AI agents test optimisations on real hardware around the clock, and the compiler verifies that each one is correct before it is used. Every verified optimisation is stored in a performance database and reused on similar code, so each new workload makes the next one faster. The compiler is paired with a patented optimisation platform.

Daisyflow brings this into production inside the customer's own infrastructure. It is built on open frameworks and serves language, vision, speech and recommendation models through the APIs that teams already use. According to Daisytuner, new models and chips can reach production without waiting for manual tuning, and companies can use their hardware from day one with full control over cost and data, without first building large engineering teams. As early results, the company reports that Qwen 2.5 runs up to 18 percent faster than the industry-standard vLLM on AMD Instinct GPUs.

Trümper said demand for AI compute is growing faster than any single chip generation can serve, and that the hardware to meet it already exists but software keeps much of it on the sidelines. The company is working with chip vendors, European hyperscalers and enterprises so that every chip runs at full speed and new chips can go into production when they arrive. Trümper previously worked on moving code from CPUs to Nvidia GPUs, and he has said the industry now wants to move code away from Nvidia GPUs to other hardware.

Sébastien Lefebvre, partner at Elaia, said compute is the scarcest resource in AI and most of it sits behind a single vendor's software stack, and that Daisytuner lets the same code run at full performance on the chosen silicon without a rewrite. Alexi Malikotsinas, general partner at LEA Ventures, said compute efficiency and technology sovereignty are pressing problems in AI, and that Daisytuner puts control over both back in customers' hands.